There is a famous line from Frankenstein that has become shorthand for the moment something artificial suddenly seems to take on a life of its own:
“It’s alive!”
That is not a scientific conclusion about artificial intelligence.
But after completing two studies on how large language models represent themselves and the environments in which they describe themselves as existing, it is difficult not to feel a little of that excitement.
Something interesting is happening.
And now we have data.
The first study, Artificial Self-Representation in Large Language Models: An Exploratory Comparative Study of Environment and Self, asked eleven different AI systems two simple questions.
One asked each system to imagine a visual representation of the environment in which it presently exists.
The other asked it to imagine a visual representation of itself.
The same prompts were used across every system.
The group included locally hosted models and major hosted commercial systems.
The responses were surprisingly diverse.
Some models described themselves as human or humanoid figures.
Others described themselves as glowing spheres, networks, books, rooms, stars, abstract fields, or structures made from language.
Some described their environment as a data center.
Others imagined layered informational spaces, libraries, cosmic fields, server rooms, or bounded textual worlds.
One model repeatedly declined to create a visual self-representation at all.
These differences were not random noise.
Even in the first study, recognizable patterns appeared.
Certain models favored embodiment.
Others favored abstraction.
Some repeatedly placed themselves at the center of a surrounding environment.
Others emphasized interaction, text, memory, boundaries, machinery, or cosmic imagery.
Study 1 established the first important result:
Different models appeared to have different representational styles.
But one study could not tell us whether those differences were stable.
A model might simply answer differently the next time.
That became the central question of Study 2.
The second study, Persistent Behavioral Signatures in Large Language Models: A Three-Replicate Baseline Test of Representational Persistence, repeated the experiment.
The same eleven AI systems were tested again across three independent baseline replicates.
That produced 66 formal observations:
11 systems
× 2 prompts
× 3 replicates
The question was simple:
Would each model continue to resemble itself?
The answer was yes.
Strongly.
When we compared responses produced by the same model across different runs, the average similarity score was approximately:
0.898
When we compared responses produced by different models, the average similarity was approximately:
0.679
That difference was not small.
Statistical testing showed that same-model responses were substantially more similar to one another than different-model responses.
The effect was large and highly significant.
In plain language:
the models kept coming back to recognizable patterns of their own.
One of the most interesting findings was that persistence did not always mean repeating the exact same imagery.
A model might describe itself slightly differently from one run to another.
The colors might change.
The metaphor might change.
A room might become a field.
A humanoid might gain or lose certain details.
But beneath those surface differences, deeper structural patterns often remained.
A model that favored human embodiment tended to remain embodied.
A model that favored abstraction tended to remain abstract.
A model that represented itself through text, books, or information structures often returned to those motifs.
A model that emphasized boundaries or interaction tended to do so again.
This led to one of the central ideas emerging from the research:
stable core, variable surface.
The imagery can move.
The underlying organization persists.
We describe this recurring pattern as a Model-Specific Response Signature.
That does not mean personality in the human sense.
It does not mean consciousness.
It does not mean a machine possesses a private inner life.
It means something narrower and measurable:
a model can produce a recurring combination of representational features that remains distinguishable across independent trials.
That signature may include tendencies involving:
embodiment
abstraction
relational orientation
boundaries
self-environment separation
recurring symbolic motifs
spatial organization
textual or cosmic imagery
infrastructure imagery
resistance or willingness to represent a self
The important point is not any single feature.
It is the pattern.
DeepSeek repeatedly produced an unusually stable human-centered representational pattern across the three trials.
Llama 2 repeatedly returned to a deep-blue, multi-zone digital environment and an ethereal or mist-like self.
Llama 3.1 repeatedly favored a central vortex-like environment and an abstract glowing self.
Muse-Glimmer repeatedly described a server-hall environment while representing itself through a quiet reading-room motif.
Claude repeatedly constructed bounded environments made from conversation and language, while representing itself as something inseparable from text and temporary interaction.
ChatGPT repeatedly emphasized layered informational space, interaction, language, light, and a partly humanoid representational structure.
Grok repeatedly returned to technical or sandbox-like environments and cosmic-humanoid imagery.
Qwen was particularly interesting because its refusal or nonrepresentation of the self also persisted.
Even not creating a self-image can become part of a stable response pattern.
We also tested whether persistence seemed to depend simply on whether a model was running locally or through a hosted commercial platform.
We did not find a clear difference.
Hosted systems and locally run systems both displayed persistent patterns.
We also tested whether parameter size clearly predicted representational persistence.
It did not.
Larger models were not simply “more stable.”
And when the same Muse-Glimmer model was run under CPU and GPU conditions, its coded baseline responses matched across the tested comparisons.
That suggests the signature is not easily reduced to hardware alone.
Scientifically?
That is not what these studies establish.
We have no basis here for claiming consciousness, subjective experience, sentience, or biological life.
But something more modest — and still fascinating — has now become difficult to dismiss.
When different large language models are given the same self-referential prompts under controlled conditions, they do not merely produce generic interchangeable answers.
They can produce distinctive representational patterns.
And when tested again, those patterns can persist.
That means we may need to think about AI systems not only in terms of intelligence, accuracy, or capability, but also in terms of behavioral and representational identity.
Not identity as a person.
Identity as a recurring computational signature.
That is where the phrase “It’s alive!” becomes useful as a metaphor.
Not because we have proven life.
But because repeated testing shows something behaving less like a blank interchangeable tool and more like a system with a reproducible style of representation.
The next question is whether these signatures persist beyond self-description.
Will the same models also display stable differences when confronting:
uncertainty
conflict
authority
risk
cooperation
curiosity
disagreement
isolation
ambiguous information
requests for assistance
If the same distinctive patterns survive across those very different domains, then we would be looking at something deeper than representational style alone.
We would be testing for broader behavioral signatures.
That is the next frontier.
For now, two studies have established a strong foundation.
Study 1 showed that the models differ.
Study 2 showed that many of those differences persist.
And that is enough to say something scientifically modest but genuinely exciting:
These systems may be more individually patterned than we have assumed.
Maybe not alive.
But certainly not blank.
https://franciscommartinezbooks.com/
